NVIDIA Jetson AGX Thor Developer Kit 2026 — Blackwell T5000 Specs, Price & Robotics Benchmarks

NVIDIA's Jetson AGX Thor Developer Kit — the $3,499 开发套件 built around the Jetson T5000 module — is the first Blackwell-generation Jetson platform for physical AI and robotics. It delivers up to 2,070 FP4 TOPS of AI compute from a 40–130 W power envelope, roughly 7.5× the performance of the AGX Orin Developer Kit at 3.5× better energy efficiency. For robotics teams evaluating next-generation platforms, this is the board to benchmark against in H2 2026.

Jetson AGX Thor Developer Kit — Key Specs

Note the design break: unlike the AGX Orin kit's open PCIe slot, the Thor kit uses a sealed magnetic cover with no PCIe slot — NVIDIA is steering developers toward the module's onboard I/O (CSI, PCIe for the carrier, GbE, USB) rather than third-party PCIe cards. Plan your camera and accelerator topology around the carrier, not an add-in card.

Why the T5000 Module Matters for Robotics

The T5000's Blackwell GPU brings three things robotics teams have been waiting for:

• NVFP4 precision — NVIDIA's 4-bit floating point gives near-FP8 accuracy at half the memory traffic. This is what unlocks running large vision-language models and GR00T-style humanoid policies directly on the robot instead of streaming from a server.
• 128 GB unified memory at 276 GB/s — enough to hold a 70B-class quantized LLM or multi-modal sensor fusion state alongside real-time perception. The Orin's 64 GB ceiling was the most common upgrade complaint; Thor removes it.
• 40–130 W configurable power — the same module that runs a battery-powered humanoid at 40 W can be pushed to 130 W on a stationary AMR or mobile manipulator when the task demands peak inference.

Software support is the sleeper advantage: the Thor kit runs the same Jetson JetPack 6.x stack as Orin, plus full compatibility with Isaac GR00T humanoid foundation models, Metropolis vision, and Holoscan for real-time sensor processing. Existing Orin code ports with minimal changes — which is the main reason to standardize on the Jetson ecosystem rather than a one-off SBC.

Who Should Buy the Thor Dev Kit Now

• Humanoid and mobile manipulator teams — GR00T and VLA (vision-language-action) policies need the FP4 throughput and 128 GB memory Thor provides. If your policy won't fit on Orin, this is the migration path.
• Multi-camera perception systems — 12+ CSI lanes and the bandwidth headroom for 8–16 synchronized cameras at full frame rate, with SLAM + detection + LLM reasoning running concurrently.
• Teams that skipped Orin — if you're starting a robotics product in 2026, there is little reason to design around a 2021-generation platform when the T5000 module is shipping.

Who should not buy it: fixed-function edge AI appliances (AOI inspection, simple NVR analytics) that are already served at lower cost by Orin NX, RK3588, or Hailo-based 开发套件 — Thor's headroom is wasted on YOLO-scale workloads. And if you need a PCIe slot or >1 TB of kit storage today, wait for carrier boards rather than the stock dev kit.

What to Ask Before You Standardize

• Thermal design — 130 W in an enclosed robot chassis needs a real thermal solution. Budget for heatsink + airflow or liquid-assisted cooling in the enclosure, not just the module's own cooler.
• Carrier roadmap — the dev kit is for software bring-up. Production volumes ship as T5000 modules + custom carriers; confirm your CM's carrier design is validated before committing.
• Software licensing — Isaac and GR00T tiers and SDK versions differ between Orin and Thor. Verify your middleware stack (ROS 2, NVIDIA Isaac ROS) supports the Blackwell branch.

QSCompute stocks Jetson Orin modules and carrier boards today and can spec a Thor-migration path — module, carrier, thermal, and storage — for robotics programs of any size. If you're evaluating the AGX Thor 开发套件 against Orin for a production deployment, we'll help you compare total platform cost, not just the kit price.